AI Diffusion in Low- and Middle-Income Countries

Low- and middle-income countries are not a delayed version of the Global North’s adoption curve. They face a different problem set: thinner fiscal space, imported standards, scarce specialised labour, and urgent sectoral needs that do not wait for frontier models to become cheap.

What makes LMIC diffusion different

Frontier discourse assumes surplus compute, surplus capital, and surplus institutional bandwidth. Most LMICs have none of these in abundance. The practical questions look more like:

Can a ministry deploy a model that works offline or under intermittent connectivity?
Who owns the fine-tuned weights trained on national health or tax data?
Does “responsible AI” guidance assume GDPR-style capacity that does not exist?
Will open weights reduce dependency — or simply shift rent extraction to cloud and chips?

These are not edge cases. They are the median case for most of the world’s population.

Agriculture: high stakes, thin margins

Agriculture remains the livelihood base for large shares of the African workforce. AI advisories — pest detection, yield forecasting, market pricing — are among the most frequently cited “leapfrog” use cases. Early deployments show a recurring pattern:

What works

Narrow, local-language tools tied to extension services and mobile money; models that tolerate low-resolution phone imagery; partnerships with farmer cooperatives rather than pure B2C apps.

What fails silently

Generic chatbots trained on temperate-climate data; pilots that die when donor funding ends; systems that require continuous API spend without a public budget line.

Diffusion here is not “percentage of farmers using ChatGPT.” It is whether public agricultural systems can afford, maintain, and govern tools that improve yields without locking data into foreign platforms.

Healthcare: capability without capacity

Diagnostic imaging support, triage chatbots, and supply-chain optimisation attract intense interest. The constraint is rarely model intelligence. It is integration: electronic records, liability regimes, clinician trust, and electricity.

A clinic that cannot keep a fridge cold will not benefit from a frontier multimodal model. The governance question follows immediately: who audits clinical AI when the regulator has five staff and no compute of its own?

Governance readiness — imported or owned?

The African Union’s Continental AI Strategy (2024) and a growing set of national strategies are real progress. Readiness, however, is not a PDF. We will track a simpler checklist over coming issues:

Procurement rules that mention AI evaluation, not only lowest bidder
Impact assessment capacity inside government, not only consultant reports
Data governance that covers training and fine-tuning, not only “privacy notices”
Local evaluation benches for language, dialect, and domain shift
Exit options — portability when a vendor changes price or geopolitics

Most countries score unevenly. Many have ethics principles. Far fewer have audit muscle. Almost none have sovereign hosting at meaningful scale.

An honest assessment

Optimism without infrastructure is marketing. Pessimism without agency is another form of extractive narrative. The LMIC story in 2026 sits between those poles: talent and need are real; capital and compute are scarce; governance is catching up unevenly; open weights help on cost and still leave dependency on chips, cloud, and standards written elsewhere.

State of AI Diffusion will keep reporting from that middle — with data when we have it, case studies when pilots survive contact with reality, and policy analysis that refuses to treat the continent as a single market.

Next quarter we will deepen the indicator set and pick one sector for a full field brief. Until then: access first, ownership next, standards with a seat at the table.